perf(rust): optimize set_bytes with direct copy_nonoverlapping

[borngraced]

Why?

All set_bytes callers follow the same pattern: save writer.len() as offset, write placeholder bytes (via write_*/skip), then call set_bytes to back-fill. Since bounds are already guaranteed by this usage pattern, the runtime checks are redundant I guess. This allows us to optimize them away while keeping a debug_assert! for safety...

What does this PR do?

Replace bounds-checked get_mut().unwrap().copy_from_slice() with direct ptr::copy_nonoverlapping.

• Does this PR introduce any public API change?
• Does this PR introduce any binary protocol compatibility change?

Benchmark

No benchmarks yet, but the improvement should be noticeable: https://godbolt.org/z/75bhPWxfW